-
Notifications
You must be signed in to change notification settings - Fork 42
/
main.py
312 lines (268 loc) · 12.9 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
from __future__ import print_function
import argparse
import os
import sys
import shutil
import json
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import transforms
from torch.autograd import Variable
import torch.backends.cudnn as cudnn
import Resnet_18
from polyvore_outfits import TripletImageLoader
from tripletnet import Tripletnet
from type_specific_network import TypeSpecificNet
# Training settings
parser = argparse.ArgumentParser(description='Fashion Compatibility Example')
parser.add_argument('--batch-size', type=int, default=256, metavar='N',
help='input batch size for training (default: 256)')
parser.add_argument('--epochs', type=int, default=10, metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--start_epoch', type=int, default=1, metavar='N',
help='number of start epoch (default: 1)')
parser.add_argument('--lr', type=float, default=5e-5, metavar='LR',
help='learning rate (default: 5e-5)')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='enables CUDA training')
parser.add_argument('--log-interval', type=int, default=250, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--resume', default='', type=str,
help='path to latest checkpoint (default: none)')
parser.add_argument('--name', default='Type_Specific_Fashion_Compatibility', type=str,
help='name of experiment')
parser.add_argument('--polyvore_split', default='nondisjoint', type=str,
help='specifies the split of the polyvore data (either disjoint or nondisjoint)')
parser.add_argument('--datadir', default='data', type=str,
help='directory of the polyvore outfits dataset (default: data)')
parser.add_argument('--test', dest='test', action='store_true', default=False,
help='To only run inference on test set')
parser.add_argument('--dim_embed', type=int, default=64, metavar='N',
help='how many dimensions in embedding (default: 64)')
parser.add_argument('--use_fc', action='store_true', default=False,
help='Use a fully connected layer to learn type specific embeddings.')
parser.add_argument('--learned', dest='learned', action='store_true', default=False,
help='To learn masks from random initialization')
parser.add_argument('--prein', dest='prein', action='store_true', default=False,
help='To initialize masks to be disjoint')
parser.add_argument('--rand_typespaces', action='store_true', default=False,
help='randomly assigns comparisons to type-specific embeddings where #comparisons < #embeddings')
parser.add_argument('--num_rand_embed', type=int, default=4, metavar='N',
help='number of random embeddings when rand_typespaces=True')
parser.add_argument('--l2_embed', dest='l2_embed', action='store_true', default=False,
help='L2 normalize the output of the type specific embeddings')
parser.add_argument('--learned_metric', dest='learned_metric', action='store_true', default=False,
help='Learn a distance metric rather than euclidean distance')
parser.add_argument('--margin', type=float, default=0.3, metavar='M',
help='margin for triplet loss (default: 0.2)')
parser.add_argument('--embed_loss', type=float, default=5e-4, metavar='M',
help='parameter for loss for embedding norm')
parser.add_argument('--mask_loss', type=float, default=5e-4, metavar='M',
help='parameter for loss for mask norm')
parser.add_argument('--vse_loss', type=float, default=5e-3, metavar='M',
help='parameter for loss for the visual-semantic embedding')
parser.add_argument('--sim_t_loss', type=float, default=5e-5, metavar='M',
help='parameter for loss for text-text similarity')
parser.add_argument('--sim_i_loss', type=float, default=5e-5, metavar='M',
help='parameter for loss for image-image similarity')
def main():
global args
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
fn = os.path.join(args.datadir, 'polyvore_outfits', 'polyvore_item_metadata.json')
meta_data = json.load(open(fn, 'r'))
text_feature_dim = 6000
kwargs = {'num_workers': 8, 'pin_memory': True} if args.cuda else {}
test_loader = torch.utils.data.DataLoader(
TripletImageLoader(args, 'test', meta_data,
transform=transforms.Compose([
transforms.Scale(112),
transforms.CenterCrop(112),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=False, **kwargs)
model = Resnet_18.resnet18(pretrained=True, embedding_size=args.dim_embed)
csn_model = TypeSpecificNet(args, model, len(test_loader.dataset.typespaces))
criterion = torch.nn.MarginRankingLoss(margin = args.margin)
tnet = Tripletnet(args, csn_model, text_feature_dim, criterion)
if args.cuda:
tnet.cuda()
train_loader = torch.utils.data.DataLoader(
TripletImageLoader(args, 'train', meta_data,
text_dim=text_feature_dim,
transform=transforms.Compose([
transforms.Scale(112),
transforms.CenterCrop(112),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
val_loader = torch.utils.data.DataLoader(
TripletImageLoader(args, 'valid', meta_data,
transform=transforms.Compose([
transforms.Scale(112),
transforms.CenterCrop(112),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=False, **kwargs)
best_acc = 0
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc = checkpoint['best_prec1']
tnet.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
if args.test:
test_acc = test(test_loader, tnet)
sys.exit()
parameters = filter(lambda p: p.requires_grad, tnet.parameters())
optimizer = optim.Adam(parameters, lr=args.lr)
n_parameters = sum([p.data.nelement() for p in tnet.parameters()])
print(' + Number of params: {}'.format(n_parameters))
for epoch in range(args.start_epoch, args.epochs + 1):
# update learning rate
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, tnet, criterion, optimizer, epoch)
# evaluate on validation set
acc = test(val_loader, tnet)
# remember best acc and save checkpoint
is_best = acc > best_acc
best_acc = max(acc, best_acc)
save_checkpoint({
'epoch': epoch + 1,
'state_dict': tnet.state_dict(),
'best_prec1': best_acc,
}, is_best)
checkpoint = torch.load('runs/%s/'%(args.name) + 'model_best.pth.tar')
tnet.load_state_dict(checkpoint['state_dict'])
test_acc = test(test_loader, tnet)
def train(train_loader, tnet, criterion, optimizer, epoch):
losses = AverageMeter()
accs = AverageMeter()
emb_norms = AverageMeter()
mask_norms = AverageMeter()
# switch to train mode
tnet.train()
for batch_idx, (img1, desc1, has_text1, img2, desc2, has_text2, img3, desc3, has_text3, condition) in enumerate(train_loader):
anchor = TrainData(img1, desc1, has_text1, condition)
close = TrainData(img2, desc2, has_text2)
far = TrainData(img3, desc3, has_text3)
# compute output
acc, loss_triplet, loss_mask, loss_embed, loss_vse, loss_sim_t, loss_sim_i = tnet(anchor, far, close)
# encorages similar text inputs (sim_t) and image inputs (sim_i) to
# embed close to each other, images operate on the general embedding
loss_sim = args.sim_t_loss * loss_sim_t + args.sim_i_loss * loss_sim_i
# cross-modal similarity regularizer on the general embedding
loss_vse_w = args.vse_loss * loss_vse
# sparsity and l2 regularizer
loss_reg = args.embed_loss * loss_embed + args.mask_loss * loss_mask
loss = loss_triplet + loss_reg
if args.vse_loss > 0:
loss += loss_vse_w
if args.sim_t_loss > 0 or args.sim_i_loss > 0:
loss += loss_sim
num_items = len(anchor)
# measure accuracy and record loss
losses.update(loss_triplet.data[0], num_items)
accs.update(acc.data[0], num_items)
emb_norms.update(loss_embed.data[0])
mask_norms.update(loss_mask.data[0])
# compute gradient and do optimizer step
optimizer.zero_grad()
if loss == loss:
loss.backward()
optimizer.step()
if batch_idx % args.log_interval == 0:
print('Train Epoch: {} [{}/{}]\t'
'Loss: {:.4f} ({:.4f}) \t'
'Acc: {:.2f}% ({:.2f}%) \t'
'Emb_Norm: {:.2f} ({:.2f})'.format(
epoch, batch_idx * num_items, len(train_loader.dataset),
losses.val, losses.avg,
100. * accs.val, 100. * accs.avg, emb_norms.val, emb_norms.avg))
def test(test_loader, tnet):
# switch to evaluation mode
tnet.eval()
embeddings = []
# for test/val data we get images only from the data loader
for batch_idx, images in enumerate(test_loader):
if args.cuda:
images = images.cuda()
images = Variable(images)
embeddings.append(tnet.embeddingnet(images).data)
embeddings = torch.cat(embeddings)
metric = tnet.metric_branch
auc = test_loader.dataset.test_compatibility(embeddings, metric)
acc = test_loader.dataset.test_fitb(embeddings, metric)
total = auc + acc
print('\n{} set: Compat AUC: {:.2f} FITB: {:.1f}\n'.format(
test_loader.dataset.split,
round(auc, 2), round(acc * 100, 1)))
return total
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
"""Saves checkpoint to disk"""
directory = "runs/%s/"%(args.name)
if not os.path.exists(directory):
os.makedirs(directory)
filename = directory + filename
torch.save(state, filename)
if is_best:
shutil.copyfile(filename, 'runs/%s/'%(args.name) + 'model_best.pth.tar')
class TrainData():
def __init__(self, images, text, has_text, conditions = None):
has_text = has_text.float()
if args.cuda:
images, text, has_text = images.cuda(), text.cuda(), has_text.cuda()
images, text, has_text = Variable(images), Variable(text), Variable(has_text)
if conditions is not None and not args.use_fc:
if args.cuda:
conditions = conditions.cuda()
conditions = Variable(conditions)
self.images = images
self.text = text
self.has_text = has_text
self.conditions = conditions
def __len__(self):
return self.images.size(0)
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def adjust_learning_rate(optimizer, epoch):
"""Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
lr = args.lr * ((1 - 0.015) ** epoch)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
if __name__ == '__main__':
main()